Research Of Language Recognition System Embedded Anchor Models And FPGA Implementation

This dissertation rely on the unit-demand project to the development of multi-channel voice processing in real time language recognition system as the overarching goal, according to the units of the actual business needs, mainstream language recognition technology on the basis of comparative analysis, focusing on the language of research-based anchor model GSV-SVM recognition system, the system back-end classification algorithm to optimize and be the FPGA implementation, the system can handle multiple voice in real time to provide a reliable guarantee. In this paper, research work and contributions are as follows:1. The analysis is based on the key technologies of the language recognition system based on the GSV-SVM, SVM kernel function is a special form of expression, an equivalent of the recognition system. The same time, the posterior probability into the scoring output of the SVM and SVM are two common classification strategy research, combined with the experimental results and the conclusion of the relevant literature, fast Pairwise Coupling method is selected as an equivalent system of back-end classification algorithm, which established in this article language recognition baseline system;2. A discriminative training algorithm is proposed based on the Model Pushing the anchor model. Baseline system for the introduction of anchor model, the algorithm through a combination of Model Pushing between the thinking of the training algorithm to construct a more languages distinguish between the anchor space, so that the GSV dimension to reduce and suppress the interference the GSV speaker information, so that the SVM training The time can be shortened. Experimental results show that the anchor Model discriminative training algorithm to improve the validity of the language recognition system;3. Combine with the above two improved algorithms for real-time processing requirements of the anchor model distinction based on the Model Pushing GSV-SVM language recognition system, a matrix combined optimization algorithm to reduce the difficulty of FPGA implementation of the system’s back-end classifier, and from therecognition performance, accuracy, resource utilization rates and real-time processing of four aspects of classifier FPGA testing and validation. The results show that the optimized FPGA implementation of the back-end classifier to complete real-time processing of up to826voice channels, and almost unanimous recognition perfonnance based on Visual Studio2010software platform to meet the accurate identification of multiple voice and real-time processing needs.